MAI-Image-2.5-Pro refines the text in the image, MAI-Voice-2-Flash sets the pace

Microsoft AI deploys MAI-Image-2.5-Pro and MAI-Voice-2-Flash. The impact of these proprietary models on the tools' computing costs and latency.

One year after initiating the development of in-house models, Microsoft AI is expanding its MAI family with two variants in public preview, positioned at opposite ends of the quality, speed, and cost curve. MAI-Image-2.5-Pro aims for maximum fidelity, targeting hero visuals, detailed editing, and precise text rendering within images, priced at one hundred and six dollars per million output image tokens. MAI-Voice-2-Flash takes the opposite approach, prioritizing volume and responsiveness, and is reportedly twice as fast as MAI-Voice-2 and thirty-two percent cheaper, at fifteen dollars per million characters.

This approach extends a fully embraced choice to train its own models on data the company presents as clean and traceable, without distillation from third-party models, rather than relying on external providers. Furthermore, the argument put forward focuses less on benchmarks than on the transition to production.

The replacement is progressing product by product. Bing Image Creator now runs entirely on in-house models. In PowerPoint, MAI-Image-2.5 handles image-to-image transformations, with a reduction in compute costs that Microsoft estimates at up to 84% compared to GPT-Image-2. Having become the default model for image editing in OneDrive, it has reportedly increased the save rate of edits by 26% and reduced latency by about a quarter. On the voice side, MAI-Voice-2-Flash powers the Dynamics 365 contact center and connects to Azure Voice Live for building voice agents.

A final component involves specialized professions. With Dragon Copilot, used by healthcare professionals, MAI-Transcribe-1.5 handles multilingual transcription in fifty-eight languages, with the company reporting, based on its internal evaluations, a halving of transcription and language identification errors for most of them.